Download Mobile Phone Enabled Social Community Extraction for Controlling

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Brucellosis wikipedia , lookup

Oesophagostomum wikipedia , lookup

Sexually transmitted infection wikipedia , lookup

Chickenpox wikipedia , lookup

Marburg virus disease wikipedia , lookup

Whooping cough wikipedia , lookup

Onchocerciasis wikipedia , lookup

Eradication of infectious diseases wikipedia , lookup

Pandemic wikipedia , lookup

Schistosomiasis wikipedia , lookup

Meningococcal disease wikipedia , lookup

Coccidioidomycosis wikipedia , lookup

African trypanosomiasis wikipedia , lookup

Leptospirosis wikipedia , lookup

Transcript
Mobile Phone Enabled Social Community
Extraction for Controlling of Disease
Propagation in Healthcare
Yanzhi Ren∗ , Jie Yang∗, Mooi Choo Chuah†, Yingying Chen∗
of ECE, Stevens Institute of Technology † Dept. of CSE, Lehigh University
Castle Point on Hudson, Hoboken, NJ 07030
Bethlehem, PA 18015
{yren2, jyang, yingying.chen}@stevens.edu
[email protected]
∗ Dept.
Abstract—New mobile phones equipped with multiple
sensors provide users with the ability to sense the world
at a microscopic level. The collected mobile sensing data
can be comprehensive enough to be mined not only for the
understanding of human behaviors but also for supporting multiple applications ranging from monitoring/tracking,
to medical, emergency and military applications. In this
work, we investigate the feasibility and effectiveness of
using human contact traces collected from mobile phones to
derive social community information to control the disease
propagation rate in the healthcare domain. Specifically,
we design a community-based framework that extracts the
dynamic social community information from human contact
based traces to make decisions on who will receive disease
alert messages and take vaccination. Our approach can be
deployed using a centralized or distributed architecture. We
have experimentally evaluated our framework via a tracedriven approach by using data sets collected from mobile
phones. The results confirmed that our approach of utilizing
mobile phone enabled dynamic community information is
more effective than existing methods, without utilizing social
community information or merely using static community
information, at reducing the propagation rate of an infectious
disease. This strongly indicates the feasibility of exploiting the
social community information derived from mobile sensing
data for supporting healthcare related applications.
I. I NTRODUCTION
The recent years have witnessed an explosion of the
usage of mobile wireless devices in our daily lives. In
particular, with the rapid deployment of sensing technology in mobile phones, the collected sensing data can
be comprehensive enough to be mined not only for the
understanding of human behaviors but also for supporting
a broad range of applications. For instance, most of the
mobile phones support the Bluetooth technology, and the
Bluetooth device-discovery software running in a mobile
phone allows it to collect information from other nearby
Bluetooth devices. It is thus convenient to exploit the
mobile phones equipped with Bluetooth technology to discover the encounter events between people such that their
social relationships can be derived and analyzed. More
importantly, the discovered social relationships can be used
to extract social communities [1], [2], which reflect close
relationships or similar behavior patterns among people,
to assist in the development of applications in various
domains, ranging from monitoring/tracking applications,
to medical, emergency and military applications.
The social community structures have been used actively in many areas including online social networks,
e.g., community detection in multi-dimensional networks
based on online social media [3], and wireless networks,
e.g., coping with the propagation of malware on smart
phones [4], and facilitating the packet forwarding in Delay
Tolerant Networks (DTNs) [5]. However, few studies
have been done in exploiting social community structures
extracted from mobile phones to control the propagation
of infectious diseases in the healthcare domain. In this
work, we focus on building a mobile phone enabled social
community based framework to reduce the rate at which
an infectious disease spreads.
In the healthcare domain, the infectious disease is a
clinically illness resulting from the presence of pathogenic
microbial agents [6], [7]. Transmission of the infectious
diseases such as SARS, bird flu and swine flu [8] can
occur when people are in close proximity. For example,
the air around a person with swine flu may contain
H1N1 virus and infect the other people close-by [9]. The
transmission of infectious diseases in public is a serious
problem related to life or death and can cause panic in
the whole society if not controlled effectively. Due to the
typical characteristic of a slow start and then exponential
propagation of the disease [10], mitigating an infectious
disease at its early stage is critical and vaccination is a
typical strategy. Because of the limited supply of vaccines
and its relatively high cost when applying to a large
population, how to efficiently distribute the vaccine and
in the meanwhile achieving the goal of effective control
of the disease propagation is an important problem.
Besides the traditional random vaccination strategy,
recent work used bridge users identified in the human
contact networks as distribution points of vaccination [11].
We are not aware of any prior work that exploits social
relationships systematically for effective vaccination such
that the propagation rate of an infectious disease can be
reduced. Since many infectious diseases propagate via
human interactions, the social communities derived from
mobile phone proximity traces in our daily lives can
be utilized to choose the set of people that need to be
vaccinated or alerted such that we can mitigate the disease
propagation more effectively and economically as opposed
to randomly choosing any person to be vaccinated or
alerted.
In light of these benefits, in this paper, we design a
social community-based method that exploits the social
relationships derived from mobile phone Bluetooth traces
to reduce the rate at which an infectious disease spreads.
Based on human’s encounter events, multiple communities
are derived and kernel structures are extracted. The community information may vary over time. Previous works
in community extraction typically find communities over
the whole trace and such static community information
is then used for making decisions, e.g., selecting the
appropriate relaying users for message dissemination [5].
However, static community extraction cannot capture the
time-varying community information present in the trace.
In our work, we propose to extract community structures at
different time periods and then merge these extracted communities to capture the dynamic community information
so as to control the disease propagation more effectively.
People who are in the same community or kernel structure are present in the close proximity more frequently
and thus may interact more with each other, whereas
those people across different communities imply fewer
interactions. We believe those people within the same
communities or kernel structures as the sick people have
higher risks to get infected, and thus should be at least
given disease alert messages and receive vaccine shots
if available. Moreover, we develop a framework which
supports two architectures, centralized and distributed,
to utilize the dynamic social community information to
control the disease spreads.
We experimentally evaluated our framework through a
trace-driven approach by using the MIT reality mining
trace [12] and the Italian trace [13]. The results showed
that our strategy is highly effective for efficient vaccination
to control disease propagation when comparing to methods
without using social relationships and schemes merely
utilizing static community information.
The rest of the paper is organized as follows. We first
put our work in the context of current research in Section II. We then present our mobile phone enabled social
community based framework in Section III. It describes
the decentralized along with the distributed system models
in our framework and the disease infection model used in
this work. We next present our dynamic social community
based scheme in Section IV. In Section V, we validate the
feasibility of our framework by using datasets collected
from mobile phones and compare with existing methods.
Finally, we conclude our work in Section VI.
rithm was introduced to improve the initial division of a
network by optimizing the number of graph edges within
and between the partitions using the greedy algorithm.
[1] developed a sociological approach called hierarchical
clustering. The idea behind this method is to develop a
measure of similarity between each pair of vertices from
the structure of the network and merge the communities
with the highest similarity. The algorithm of Girvan and
Newman [2] divided the network by iteratively removal
of the edges. The betweenness metric is a centrality
measure of a vertex within a graph. Vertices that occur
on many shortest paths between other vertices have higher
betweenness than those that do not. [15] further analyzed
the computational cost of the betweenness metric in social
networking applications.
Another important metric in community detection is
modularity Q as described in [3], [16] where a larger modularity indicates more frequent within-group interaction. In
general, one aims to find a community structure such that
Q is maximized. On the other hand, instead of relying on
a centralized server, [17] proposed distributed community
detection, which makes mobile devices sense and detect
their own local communities.
The active development of group discovery and community detection provides promising techniques for applying social relationships to support various application
domains. [3] performed online group discovery in multidimensional networks obtained from various social media
(e.g., YouTube and Flickr). [18], [19] conducted social
network analysis in Delay Tolerant Networks by utilizing
betweenness and similarity metrics. Moreover, the social
community structures were utilized to cope with the propagation of malware on smartphones in mobile networks
as proposed in [4], and [10] developed a social network
based patching scheme for effectively limiting the spread
of MMS and SMS based worms in cellular networks.
However, little work has been done in applying social
community structures to effectively control the disease
propagation through vaccination in the healthcare domain.
[20] studied the relationships between the voluntary vaccination and the transmission of a vaccine-preventable
infection. It pointed out that the propagation of the disease
is related with the neighborhood size. [21] proposed to
study contact networks, where each person’s role in a
population is treated as distinct, i.e., a heterogeneous
population. It suggested that by restricting the contacts
themselves, one can also limit disease spread effectively.
This would correspond to deleting edges in the modeled
contact network.
[11] further considered the modeling of the disease
spread over populations. It proposed that vaccinating the
groups of more sociable persons can prevent a larger number of infectious than if administering the same number
of vaccinations to random members of the population.
None of these works have systematically investigated the
II. R ELATED W ORK
Group discovery and community detection have been
an active research area. In [14], the Kernighan-Lin algo2
Susceptible
(without alert)
Susceptible
(with alert)
Susceptible
(without alert)
Chosen by the sick
person at each encounter
pre u pim
Immunized
Chosen by the sick
person at each encounter
Infective
(a) Centralized architecture
Fig. 1.
pre u (1 pim )
pin
Chosen by the server
Infective
Susceptible
(with alert)
pin u pal
pin u pal
pre u (1 pim )
pin
Chosen by the server
pre u pim
Immunized
(b) Distributed architecture
Epidemic infection model used in our framework.
effectiveness of exploiting social community structures for
efficient vaccination such that the propagation rate of an
infectious disease is reduced. Our work is novel in that
we extracted dynamic social community information by
leveraging the contact traces derived from mobile phones
and proposed a community based framework for control
of disease propagation.
analysis. Our framework will utilize the existing infrastructure in cellular networks. We assume the users are
subscribed to the cellular data plan and recorded encounter events (which include discovered device IDs and
timestamps) will be periodically sent back to a back-end
server authorized by the service provider. The dynamic
community extraction mechanism is run by the server. The
detailed description of our dynamic community extraction
approach is presented in Section IV. Moreover, the extracted community information will be stored at the server
and updated from time to time.
2) Centralized vs. Distributed Architecture: We design
two types of messages that a user may receive: vaccination
and alert. A user who receives a vaccination message
should go to obtain a vaccine shot, whereas a user receiving a alert message should take precautions as directed.
We assume that all the users who have been notified
will take the necessary recommendated actions. In our
framework, vaccine shots of an infectious disease only
have limited supplies and are more costly comparing to
alert messages. The number of alert messages for each
disease can be either controlled or unlimited. To protect the
useres’ privacy, such messages will be sent anonymously
so that the receivers do not know who the senders are. The
full discussion of the privacy issue is out of the scope of
this paper and will be included in our future work.
When actions need to be taken for an infectious disease,
in the centralized architecture, the server will decide on
who will receive vaccination messages and who will
receive alert messages respectively based on the extracted
social communities stored in its database. Then the server
will send out each message to corresponding users.
On the other hand, in the distributed architecture, each
user who has already been infected by the disease will
download the community information related to his device
ID from the server. We note that although the user based
community information only contains a partial view of
the whole community information stored in the server, the
memory requirement of storing such partial community
information is much less than that of storing the whole
community information. This makes it applicable to store
the partial community view in individual mobile devices.
Based on the downloaded social community information,
III. F RAMEWORK OVERVIEW
In this section, we first provide the system model for
our mobile phone enabled disease control framework, and
present descriptions of both the centralized and distributed
architecture in our framework. We envision this framework
can be implemented by any State Department of Health
through the coordination of the Centers for Disease Control and Prevention (CDC). For example, during the 2009
spreading period of the pandemic influenza A (H1N1)
virus, every state in US is required to report the number of
infected patients to the CDC. The available vaccines are
then allocated appropriately by the CDC to the different
states [22].
We then present the infection model used in our work
and provide an analysis on the state transitions in our
model. Without loss of generality, we do not consider the
differences between users and assume that all the users
follow the same infection model.
A. System Model
1) Uncovering Human Social Relationships from Contact Traces via Mobile Phones: Instead of random vaccine
distribution, targeting vaccination to a group of people
with higher risk of infection can provide more effective
control of an infectious disease propagation. Traditionally,
scientists and doctors have to rely on social relationships
derived via manually recorded daily activities from human
subjects [12]. However, this approach is tedious, errorprone as the human subjects may forget to perform recording from time to time, and can be out of date. In this work,
we consider extracting social community information from
human contact traces collected by mobile phones.
The Bluetooth enabled device-discovery process is simple and automatic, and thus is suitable for recording
encounter events between people for social relationship
3
Notation
pin
pin × pal
pre
pre × pim
pre × (1 − pim )
Np
Description
Disease infection probability
Disease infection probability
with alert messages
Recovery probability
after the recovery cycle
Probability of recovery with immunization
after the recovery cycle
Probability from infective to susceptible
after the recovery cycle
The length of the disease recovery cycle
downloaded from the server and the next set of people he
will encounter.
(3) From Susceptible (without alert) to Infective: If a
user encounters a sick person, the user has a probability
pin of being infected.
(4) From Susceptible (with alert) to Infective: Since precautions will be taken for users at the Susceptible (with
alert) state, the probability of such a user get infected
when he encounters a sick person is reduced to pin × pal .
(5) From Infective to Immunized: An infected user
moves to the Immunized state after he has recovered from
the disease with immunity. This happens with a probability
of pre × pim every recovery cycle (set to Np days).
(6) From Infective to Susceptible (without alert): An infected user moves to the Susceptible (without alert) state if
he has recovered from the disease, but not immunized from
being infected again. Based on item (5), this probability
should be pre × (1 − pim ) for every recovery cycle.
TABLE I
N OTATIONS USED IN THE INFECTION MODEL .
each sick user will then decide on whether to send out a
vaccination message or a alert message when he encounters the next person, i.e., a new person discovered by the
sick person’s mobile device.
The distributed approach is a challenging architecture
as it requires peer-to-peer communications over an ad hoc
wireless network. Moreover, in our future work, we plan
to develop distributed social community extraction scheme
so that each user has the capability to derive his local
community and does not rely on the centralized server.
C. Analysis of State Transitions
The epidemic infection model includes rules on how the
members of users transit from one state to another. Directly
analyzing the state transitions within the social community
model is prohibitive. Thus, we analyze the state transitions
based on users’ encounter events. Because we believe
that this analysis can provide us with an upper bound of
the final infected ratio for our dynamic community based
system since the contact rate is assumed to be constant
in the analytical model, although in real system this rate
decreases as more people gets infected or vaccinated.
We assume that the total number of users Ntotal will
not change and users only transit from one state to another
state. Let S1 (t) and S2 (t) denote the number of users in
the state Susceptible (without alert) and Susceptible (with
alert), respectively. I(t) denotes the number of users in
the state Infective and R(t) denotes the number of users
Immunized. Time t is taken as a variable here. We assume
that the encounters between infected users and other ones
occur at an average rate of β, which is often referred to as
the contact rate. The state transitions can be represented
through the following equations:
B. Infection Model
In our framework, we extend the standard epidemic SIR
model [23] to four states: susceptible without alert, susceptible with alert, infective and immunized. Susceptible
means that a user can be infected by the disease. When a
user is susceptible, he can be at either susceptible without
alert or susceptible with alert. When a user is infected, he
goes to the infective state and he can infect other people
that he encounters. A user may go to the immunized state
only when he is either vaccinated or has recovered from
the disease with immunity.
The notations used in the infection model and across
the paper are summarized in Table I. We depicted the state
transition diagram of our infection model in Figure 1 and
list all possible state transitions as follows:
(1) From Susceptible (without alert) to Immunized: In
the centralized architecture, the server will choose the
users who should receive the vaccination messages based
on the extracted social community information and send
the vaccination messages to these users directly. Whereas
in the distributed architecture, the user who is already
infected will decide on who should receive the vaccination
messages based on the social community information
downloaded from the server and the next set of people
he will encounter.
(2) From Susceptible (without alert) to Susceptible (with
alert): Likewise, in the centralized architecture, the server
chooses the users who should receive the alert messages
based on the extracted social community information
and sends the alert messages to these users directly.
Whereas in the distributed architecture, the user who is
already infected will determine who should receive the
alert messages based on the social community information
S1 (t+∆t)−S1 (t)
=−β·I(t)·S1 (t)·(1−pva )−β·I(t)·S1 (t)·pva
∆t
+pre ·(1−pim )·I(t)·δ(t−Np ·N )
S2 (t+∆t)−S2 (t)
=β·I(t)·S1 (t)·(1−pva )−β·I(t)·S2 ·(t)·pin ·pal
∆t
I(t+∆t)−I(t)
∆t
(2)
=β·I(t)·S2 (t)·pin ·pal −pre ·pim ·I(t)·δ(t−Np ·N )
−pre ·(1−pim )I(t)δ(t−Np ·N )
R(t+∆t)−R(t)
=β·I(t)·S1 (t)·pva +pre ·pim ·I(t)·δ(t−Np ·N )
∆t
4
(1)
(3)
(4)
1
1
Susceptible(without alert)
Susceptible (with alert)
Infective
Immunized
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.8
0.7
For the community information, we define two types
of social clusters to represent different levels of social
relationships: one is refereed to as community and the other
is referred to as kernel structure. The people within the
same community meet frequently with one another, while
the kernel structure aims to capture a subset of people
on top of the community structures that have even higher
encounter frequency. Instead of using static community
information derived from the whole trace, we propose
an approach called dividing and merging, where dynamic
community information is utilized since people may belong to different social communities at various times, and
communities may appear or disappear in different time
periods.
Flow Overview. Our dynamic community and kernel
extraction approach is illustrated in Figure 3. First, mobile
phones with Bluetooth capability record user encounter
events. The recorded human encounter events are divided
into multiple trace files based on each time window. We
note that the length of the time window is adjustable (e.g.,
the length of the time window can be one day).
From each contact trace file, two contact graphs are
constructed. One will be used for extracting community
and the other for extracting kernel structure. Hierarchical
clustering method can be one of the options to extract
both community and kernel structures. The extracted community and kernel structures learnt for the current time
period are then merged with the existing community and
kernel structures that our system maintains. The combined
community and kernel structures will be used to make decisions on who to send the vaccination and alert messages
when the next request coming from the server or from an
individual user based on the specific architecture used in
our framework.
In following subsections, we first describe how we
construct contact graphs and our dividing strategy for the
construction of community and kernel structures. Then, we
describe how we merge the newly learnt community and
kernel structures from different trace files with the existing
community information.
0.6
0.5
0.4
0.3
0.2
0.1
0
0
2
4
6
8
0
0
10
2
4
Days
6
8
10
Days
(a) β = 0.01
(b) β = 0.06
1
1
Susceptible(without alert)
Susceptible (with alert)
Infective
Immunized
0.8
Susceptible(without alert)
Susceptible (with alert)
Infective
Immunized
0.9
Ratio of users at each state
0.9
Ratio of users at each state
IV. DYNAMIC E XTRACTION OF C OMMUNITY
I NFORMATION
Susceptible(without alert)
Susceptible (with alert)
Infective
Immunized
0.9
Ratio of users at each state
Ratio of users at each state
0.9
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
2
4
6
8
10
0
0
2
Days
(c) Np = 1 day, β = 0.02
4
6
8
10
Days
(d) Np = 4 days, β = 0.02
Fig. 2. Ratio of the users at different states when varying recovery cycle
and contact rate.
In the equations above, the N = 1, 2, 3, ... and the function
δ(t) is defined as:
(
1 if x = 0;
δ(x) =
(5)
0 otherwise.
For the above equations, we also have:
S1 (t) + S2 (t) + I(t) + R(t) = Ntotal .
(6)
D. Analytical Results
Based on the above state transition analysis, we vary the
parameters in the infection model to provide an illustration
of the number of users at each state as a function of time.
We first vary the contact rate β to study its impact: we
consider a contact rate of 0.06 and 0.01. Based on the
results depicted in Figure 2 (a) and (b), we concluded that
the infected ratio remained low for the whole duration
when users meet less frequently, while the number of
infected people increases significantly if users meet more
frequently. This indicates that our infection model can
precisely capture the propagation trend of an infectious
disease.
We then set the contact rate β = 0.02 based on the
mean value observed from the MIT reality trace [12] and
examine the number of users at different states when the
initial ratio of the sick person is set to 0.15. Figure 2
(c) and (d) plotted the results for 10 testing days when
recovery cycle is 1 day and 4 days respectively. We
observed that the ratio of the infected persons increases
as the time moves along. In particular, it increases to 0.2
and 0.45 for 1 day and 4 days recovery cycle at the end
of the testing period.
A. Dividing Based Contact Graph Construction
The whole contact trace is divided into multiple trace
files. We assume that each trace file consists of recorded
encounter events that happened during the time period
[Ti , Ti+1 ]. Each entry in such a trace is a record of one
encounter event between two mobile phones: including
the starting and ending time of the contact as well as
unique IDs of the mobile phones. We also assume that
the same person carries the mobile phone for the duration
of the trace. Based on this information, a contact graph
G = (V, E) can be derived, which consists of a vertex
set V and an edge set E. Each vertex u ∈ V denotes a
5
Dividing
Process
T1
Merging Process
Community
Structures
CS1
Modularity Q
Hierarchical
Clustering
"
Mobile Phone
Traces
During Time
Period
Construct
Contact Graphs
(Community
Level )
Fig. 3.
"
TD
"
"
Mobile Phone
Traces
During Time
Period
Construct
Contact Graphs
(Community
Level )
Community Merge
Kernel
Structures
KS1
Construct
Contact Graphs
(Kernel Structure
Level )
Kernel Structure
Merge
Community
Structures
Modularity Q
CS D
Hierarchical
Clustering
Kernel
Structures
Construct
Contact Graphs
(Kernel Structure
Level )
KS D
Community
Structure
Output
Kernel
Structure
Output
Flow overview: components to extract community information.
person, while each edge e(u, v) denotes that person u has
contacted person v for at least W times. The weight t(u, v)
denotes how frequent the two persons u and v meet during
[Ti , Ti+1 ]. We use the number of times that the two persons
have encountered with each other as the weight because
the people who encounter with each other frequently tend
to have closer relationships or similar social behaviors (e.g.
riding on the same train to go to work each morning).
ones. We next describe our community merging technique.
We note that the same technique is applied to merge kernel
structures.
We assume that we have D time windows. We have
constructed one contact graph from each time period
and we assume these are non-overlapping time periods:
[T0 , T1 ], [T1 , T2 ],..., [TD−1 , TD ] with G1 = (V1 , E1 ),
G2 = (V2 , E2 ),..., GD = (VD , ED ). The communities are
extracted from each contact graph Gi = (Vi , Ei ) by using
the hierarchical clustering algorithm and the modularity Q.
Let Si represents the set of communities found for time
window i. Thus, we have S1 , S2 , ..., SD . Each Si contains
a set of vertices Ai . Each Ai has been divided into ki
communities, which are represented as follows:
B. Community & Kernel Structures Extraction From Contact Graphs
For each trace file, we construct two contact graphs:
one with W = w1 and the other with W = w2 where
w2 > w1 . Clusters extracted using the first contact graph
are referred to as communities, whereas clusters extracted
from the second one are referred to as kernel structures.
An illustration of community and kernel structures is
shown in Figure 4. Two communities can be extracted
using a contact graph. Community 1 consists of users {A,
B, C, D, H} while Community 2 consists of users {E,
F, G, I}. The communities are shown as dotted circles.
Moreover, by using a higher W , three sets of users, namely
{A, B,C}, {E, F , G} and {D, H} are found having
higher encounter frequencies within the same set than the
rest of users. Thus, they form 3 different kernel structures
(shown with solid circles).
For scalability, it is important that an efficient algorithm
is used to partition the contact graph G = (V, E) into
separate clusters. In this paper, we use a simple, yet
effective partition algorithm called hierarchical clustering [24]. Further, to verify whether a particular division is
meaningful or not, we use the modularity metric, Q [24].
This metric has often been used by researchers in previous
studies to measure how good a partition is. A larger Q
value indicates a better partition of the users.
Ai = A1i ∪ A2i ∪ ... ∪ Aki i
(7)
We compare each community in Si with all the communities discovered in Si+1 to see if a community in Si
satisfies one of the following conditions:
• It is part of a bigger community in Si+1 and hence
can be removed.
• It can be merged with one community in Si+1 using
the community merge operation for two communities
Aji and Ali+1 under an adjustable threshold τ :
|Aji ∩ Ali+1 |
M ax(|Aji |, |Ali+1 |)
>τ
(8)
It is a superset of a community Aji+1 in Si+1 , then
Aji+1 is removed from set Si+1 .
At the end of this operation, the two sets Si and Si+1 are
0
unioned to form a new Si+2 , which will merged with Si+2
in the next round of comparison. The merging process
iterates through D time windows.
•
D. Using Extracted Community Information in Disease
Propagation Control
C. Merging Community Information Extracted over Different Time Periods
Based on our community information extraction strategy, people that belong to the same kernel structures have
a higher encounter frequency. Thus, an infectious disease
has a higher probability to spread among these group of
Recall that the social community information may
change with time: some communities may merge, some
may disappear, and others may be divided into smaller
6
people if one person is infected already. Similarly, those
in the same community as a sick person are also more
susceptible to be infected by the disease. However, the
probability for the disease to spread across two disjoint
communities is low because people in such communities
contact less frequently. We note that one person can belong
to multiple communities and kernels. Let Vs represent the
set of sick persons for a particular infectious disease. We
define the susceptible persons who are in the same kernels
as the sick people as Vk , while those susceptible persons
who are in the same communities but are not in the same
kernels as those sick people as Vc .
Because of the limited supply and relatively high cost of
vaccines, an appropriate decision on efficient vaccination
is that the vaccine shots and the alert messages should
be given to those people who have higher risk of being
infected by the disease. Thus, by utilizing the community
information, the people in Vk should have higher priority
to receive vaccination or alert messages than those in Vc .
We further define the importance of a person by the weight
when there are total M number of extracted communities
(or kernel structures):
Definition 1. The weight W (v, S) of a person v in the
community (or kernel structure) set S = V1 , V2 , ..., VM is
defined as the total number of people in the community
(or
P
kernel structure) that v belongs to: W (v, S) = (|Vj |−1)
for all Vj which satisfies v ∈ Vj .
We further return the top K user list based on the
following function:
Definition 2. The T OP (V, K) is defined as the function
which can return a Top-K ranked list of the persons in V
based on their weights W .
Our goal is to find two optimum sets of people, one
for receiving vaccination messages for vaccine shots, and
the other for receiving alert messages, such that we can
keep the infection rate low and effectively control the
propagation of the disease. We next describe how these
two sets of users are selected in our community-based
framework.
Centralized Community Based Algorithm. As described in Section III, the server will decide who to receive
vaccination or alert messages. The flow of the centralized
community based algorithm is as follows:
• The kernel structures Vk are considered first and the
weight of each person in Vk reflects the priority. The
function T OP (Vk , K) is called to produce the top K
user list Lk , where K is determined by the number
of available vaccination or alert messages.
• If there are remaining vaccination and alert messages after considering all the people in Vk , then
the community structures Vc are considered and the
weight of each person in Vc reflects the priority. The
T OP (Vc , K) function is called to return the top K
user list Lc , where K will be set to the remaining
value of vaccination or alert messages.
B
K1
23
C1
A
8
D
26
19
25
9
K2
H
C
3
3
2
2
G
C2
Fig. 4.
22
17
I 9
E
28
K3
F
An example of social community.
In the case that the number of these messages is
largerSthan the number of total susceptible persons
in Vc Vk , the remainder messages are held till the
next update of the community information as new
persons may appear.
We note that for every round of calculation, choosing
the candidates to send out the vaccination messages will
take higher priority than alert messages.
Distributed Community Based Algorithm. The sick
person i downloads his community set Vci and kernel
structure set Vki information from the server. In the meantime, the sick person will check the availability of the
vaccination and alert messages with the server. We note
that the community and kernel structure sets may also
be possibly determined by users themselves, e.g., [17].
This issue is further explored in our future work. The sick
person then performs the following:
i
• The sick person sets the persons who are in Vk as
candidate user list for sending out the vaccination
messages.
i
• The sick person sets the persons who are in Vc as
candidate user list for sending out the alert messages.
• The sick person sends out corresponding messages
when he encounters another user who belongs to the
above candidate lists. We note that if there are not
enough vaccination messages for the people in Vki ,
the alert messages will be sent instead.
•
V. P ERFORMANCE E VALUATION
In this section, we first describe our simulation methodology and present three existing methods for vaccination
distribution. We then present the performance of our social
community based methods by comparing to the existing
techniques.
A. Simulation Methodology
We implemented our framework in a home-grown tracedriven simulator. We used two human contact-based traces,
namely the MIT reality [12] and Italian [13] traces. Both
traces were collected using smart phones equipped with
7
1
Ratio of infected persons
0.8
0.7
0.6
0.5
0.4
0.3
0.9
0.8
Ratio of infected persons
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
0.7
0.4
0.3
0.1
0.1
4
Recovery Cycle (days)
8
(a) Scenario 1: Initial sick ratio: 0.03
0.8
0.5
0.2
1
0.9
0.6
0.2
0
1
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
Ratio of infected persons
1
0.9
0
0.7
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
0.6
0.5
0.4
0.3
0.2
0.1
1
4
Recovery Cycle (days)
8
(b) Scenario 2: Initial sick ratio: 0.15
0
1
4
Recovery Cycle (days)
8
(c)Scenario 3: Initial sick ratio: 0.3
Fig. 5. MIT traces: Performance comparison under different recovery cycle Np when there are 15 vaccines and 30 alert messages with pin = 0.5,
pal = 0.7, pre = 0.1, pim = 0.2, pva = 0.3.
bluetooth devices. Each trace contains information about
the IDs of the Bluetooth devices which are within the
transmission range of each other, and the starting and
ending times of their encounter. The MIT traces were
collected from smart phones carried by 97 participants in
an university environment. We used the first 20 days of
the MIT traces which contains encounter events from 71
people. In particular, we used the first half of the trace
(i.e., 10 days) as training data to extract the communities
and kernel structures, and the second half trace as the
testing data to evaluate our approach. In the Italian traces,
there are 44 people who carried the smart phones and
the experiment lasted for 19 days. Similarly, we used the
first 9 days of the trace as training data to extract the
communities and kernel structures, and the remaining 10
days to evaluate our approach. We conducted extensive
experiments on these two sets of traces by varying different
parameters in our epidemiology infection model. Due to
the space limit, we only present a subset of the results in
the following subsections.
thus is static. Each susceptible person is ranked based on
the betweenness metric, and then the persons with higher
ranks will be chosen to receive vaccination messages while
the persons with lower ranks will be chosen to receive alert
messages. This approach has an advantage of obtaining a
set of more active users from the population than Random
Distribution.
C. Effectiveness of Disease Propagation Control
In the first set of experiments, we evaluate the effectiveness of our social community based methods in terms
of the final ratio of infected persons at the end of our test
by comparing to existing methods of Random Distribution
and Encounter-based. We use the MIT traces and vary both
the recovery cycle Np and the initial sick ratio. Figure 5
presents the final ratio of the infected persons versus the
recovery cycle. The No vaccine is plotted as a baseline
case.
The key observation is that our proposed community
based methods, both centralized and distributed, consistently achieve a lower infection ratio than Random Distribution and Encounter-based methods for each initial sick
ratio and each recovery cycle. This observation is also
inline with our analytical results depicted in Figure 2 (a)
and (b). This is very encouraging since the persons chosen
by our community-based methods to have vaccination or
alert messages interact more frequently with each other.
Consequently, the proposed community based methods
can control the disease propagation more effectively than
other methods. We also found that the final infection ratio
increases when the initial infection ratio or recover cycle
increases for all the methods.
Moreover, we found that the performance of our centralized community based method is better than its distributed
version. This is because in the centralized approach, the
distribution of vaccination or alert messages is based on
the complete picture of the social community information,
instead of the partial local community information in the
distributed case.
B. Existing Methods
We compare our social-community based approach to
the following three existing methods for efficient vaccine
distribution to achieve effective disease propagation control.
Random Distribution Method. This is the most
straight forward method. In this method, the server will
randomly choose the users to receive the vaccination and
alert messages.
Encounter-based Method. This method involves message distribution based on the encounter of mobile phones.
We apply the scheme in [25] and let the sick user to send
out messages when it encounters a susceptible person. In
our simulation, once the sick person encounters with a
susceptible person, the vaccination message is sent with
the probability pva , while the alert message is sent with
the probability 1 − pva .
Betweenness-based Community Method. This method
is an improvement over Random Distribution. The concept
of the betweenness [2] is used to identify a set of key users
that act as bridge users in a contact network [11]. The
extraction of bridge users is based on the whole trace and
D. Impact of the Number of vaccination and alert Messages
Next, we change the available number of the vaccination and alert messages under different initial ratios
8
0.7
0.8
Ratio of infected persons
Ratio of infected persons
0.8
0.6
0.5
0.4
0.3
0.7
0.9
0.6
0.5
0.4
0.3
0.8
0.5
0.4
0.3
0.2
0.1
0.1
0.1
0.1
0.15
0.2
Initial ratio of sick persons
0.25
0
0.3
(a) vaccine: 15; alert: 30
0.9
0.9
0.8
0.8
Ratio of infected persons
1
0.7
0.5
0.4
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
0.3
0.2
0.15
0.2
Initial ratio of sick persons
0.25
0
0.3
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
0.5
0.4
0.1
0.15
0.2
Initial ratio of sick persons
0.25
(c) vaccine: 35; alert: 30
0.3
0
0.05
0.1
0.15
0.2
Initial ratio of sick persons
0.7
0.6
0.5
0.4
0.3
0.1
15
35
Number of available vaccines
0
15
35
Number of available vaccines
(b) Scenario 2: Initial sick ratio: 0.24
for different methods when compared to the No vaccine
scenario, whereas the x-axis shows the number of vaccines
used in our community-based methods and Betweenness.
The results clearly showed that our community based approach outperforms the static betweenness based method,
especially when there are fewer number of vaccines is
available (i.e., the case with 15 vaccines) for both initial
sick ratios of 0.09 and 0.24. This is because Betweenness
selects the relaying users for message dissemination from
static community information. However, static community
information cannot capture the time-varying community
information present in the trace. On the contrary, our
community based method uses dynamic community information produced from our community extraction process,
and hence it can control the disease propagation more
effectively.
0.3
0.1
0.05
Community based (distributed)
Community based (centralized)
Betweenness
Fig. 7. MIT traces: Comparison of community based methods and
the static betweenness method under different number of vaccines when
there are 30 alert messages with pin = 0.5, pal = 0.7, pre = 0.1,
pim = 0.2, pva = 0.3, Np = 4 days.
0.7
0.6
0.8
0.2
(a) Scenario 1: Initial sick ratio: 0.09
0.2
0.1
0
0.1
(b) vaccine: 15; alert: unlimited
1
0.6
0.05
0.9
0.6
0.2
0.05
1
Community based (distributed)
Community based (centralized)
Betweenness
0.7
0.2
0
Ratio of infected persons
1
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
0.9
Percentage of improvement
1
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
Percentage of improvement
1
0.9
0.25
0.3
(d) vaccine: 35; alert: unlimited
Fig. 6. MIT traces: Performance comparison under different number of
vaccination and alert messages with pin = 0.5, pal = 0.7, pre = 0.1,
pim = 0.2, pva = 0.3, Np = 4 days.
of infected persons. The results from MIT traces are
presented. Figure 6(a) and (b) depicted the results of 30
alert messages and unlimited alert messages respectively,
when the available number of vaccines is 15, which is
about 20% of the total number of people in the experiment.
While Figure 6(c) and (d) presented the results of 30 alert
messages and unlimited alert messages respectively, when
the available number of vaccines is 35, which is about 50%
of the total number of people in the experiment. Again,
we observed that our proposed community based method
can achieve a much lower final infection ratio than the
Random Distribution and Encounter-based methods under
different number of vaccination and alert messages.
Furthermore, we found that there is an increasing trend
of the infection ratio as we increase the initial ratio of
sick persons. However, the final infection ratio decreases
as the number of alert messages increases from 30 to
unlimited. This is consistent with our expectation: more
alert messages allow more people to take the necessary
precautions, which reduce their chances of being infected,
and hence reducing the number of total infected people.
In addition, comparing the results in Figure 6 under
different vaccine numbers, we found that the performance
difference between our proposed community based methods and other methods is smaller when increasing the
vaccine number from 15 to 35. This further indicates that
our proposed approach is more effective when the supply
of vaccine is limited.
F. Results from Italian Traces
Finally, we present our study using the Italian traces.
Due to the space limitation, only the key results are
presented. We changed the available number of the vaccination and alert messages using different initial ratios of
infected persons and examined the final ratio of infected
persons. The results depicted in Figure 8 exhibit the
same trend as we observed when using MIT traces: our
social community based methods outperforms the existing
methods.
Comparing these results with those obtained from MIT
traces in Figure 6, we further observed that the performance difference between our proposed community based
methods and the existing methods is relatively smaller in
the Italian traces. We found that this is because there is a
large group of people who always encounter one another
in the Italian traces and hence they belong to the same
kernel structure and community. Thus, our community
based methods can not distinguish further among these
people when they attempt to choose a subset of appropriate
people for receiving the vaccination or alert messages.
E. Comparison with the Static Betweenness Method
VI. C ONCLUSION
In Figure 7, we compare the results obtained from
our community-based approach and the static betweenness
based community method (shortened as Betweenness). The
y-axis shows the improvement in the final infection ratio
In this paper, we proposed a mobile phone enabled community based disease control framework, which utilizes
human social relationship information to reduce the rate
9
1
0.9
0.8
0.8
0.7
0.6
0.5
0.4
0.3
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
0.2
0.1
0
0.05
0.1
0.15
0.2
Initial ratio of sick persons
0.25
Ratio of infected persons
Ratio of infected persons
1
0.9
0.6
0.5
0.4
0.3
0
1
0.9
0.9
0.6
0.5
0.4
0.3
0.2
0.15
0.2
Initial ratio of sick persons
0.25
0.3
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1
0.8
Ratio of infected persons
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
0.7
0.05
(b) vaccine: 9; alert: unlimited
1
0.8
No vaccine
Random Distribution
Encounter Based
Community Based (distributed)
Community Based (centralized)
0.1
(a) vaccine: 9; alert: 18
Ratio of infected persons
0.7
0.2
0.3
[2] M. Girvan and M. E. J. Newman, “Community structure in social and biological networks,” in in Proceedings of the National
Academy of Sciences of the United States of America, June 2002.
[3] L. Tang, X. Wang, and H. Liu, “Uncovering groups via heterogeneous interaction analysis,” in in Proceedings of IEEE International
Conference on Data Mining(ICDM), 2009.
[4] F. Li, Y. Yang, and J.Wu, “Cpmc: An efficient proximity malware
coping scheme in smartphone-based mobile networks,” in in Proceedings of IEEE Infocom, 2010.
[5] F. Li and J. Wu, “Localcom: A community-based epidemic forwarding scheme in disruption-tolerant networks,” in in Proceedings of
of IEEE Communications Society Conference on Sensor, Mesh and
Ad Hoc Communications and Networks(SECON), 2009.
[6] K. J. Rothman, S. Greenland, and T. L. Lash, Modern Epidemiology. lippincott williams and wilkins, 2008.
[7] R.M.Anderson and R.M.May, Infectious Diseases of Humans:
Dynamics and Control. Oxford University Press, 1992.
[8] , “Introduction to influenza,” Nov. 2010. [Online]. Available:
http://en.wikipedia.org/wiki/Influenza
[9] , “Public health fact sheet,” Nov. 2009. [Online]. Available:
www.mass.gov/Eeohhs2/docs/dph/cdc/factsheets/swine_flu.pdf
[10] Z. Zhu, G. Cao, S. Zhu, S. Ranjan, and A. Nucci, “A social network
based patching scheme for worm containment in cellular networks,”
in in Proceedings of IEEE Infocom, 2009.
[11] S. Huang, “Probabilistic model checking of disease spread and
prevention,” in in Scholarly Paper for the Degree of Masters in
University of Maryland, 2009.
[12] N. Eagle and A. Pentland, “Reality mining: Sensing complex social
systems,” In Personal and Ubiquitous Computing, vol. 10, no. 4,
2005.
[13] S. Gaito, E. Pagani, and G. P. Rossi, “Opportunistic forwarding
in workplaces,” in in Proceedings of the Second ACM SIGCOMM
Workshop on Online Social Networks, 2009.
[14] B.W.Kernighan and S.Lin, “An efficient heuristic procedure for
partitioning graphs,” in Bell System Technical Journal, vol. 49, pp.
291–307, 1970.
[15] U. Brandes, “A faster algorithm for betweenness centrality,” In
Journal of Mathematical Sociology, vol. 25, pp. 163–177, 2001.
[16] M. E. J. Newman and M. Girvan, “Finding and evaluating community structure in networks,” in Physical Review E, vol. 69, no. 2,
2004.
[17] P. Hui, E. Yoneki, S. yan Chan, and J. Crowcroft, “Distributed
community detection in delay tolerant networks,” in in Proceedings
of Sigcomm Workshop MobiArch, 2007.
[18] E. Daly and M. Haahr, “Social network analysis for routing in
disconnected delay-tolerant manets,” in Proceedings of the ACM
International Symposium on Mobile Ad-Hoc Networking and Computing (MobiHoc), 2007.
[19] M. Chuah and A. Coman, “Identifying connectors and communities: Understanding their impacts on the performance of a dtn
publish/subscribe system,” in in Proceedings of International Conference on Computational Science and Engineering, 2009.
[20] A. Perisic and C. T. Bauch, “Social contact networks and disease
eradicability under voluntary vaccination,” In PLoS Computational
Biology, vol. 5, p. e1000280, 2009.
[21] L. A. Myers, “Contact network epidemiology: Bond percolation
applied to infectious disease prediction and control,” American
Mathematical Society, vol. 44, no. 1, pp. 63–86, 2007.
[22] , “Morbidity and mortality weekly report,” Mar. 2010. [Online]. Available: http://www.cdc.gov/mmwr/preview/mmwrhtml/
mm5914a3.htm?s_cid=mm5914a3_e
[23] H. Yuan, G. Chen, J. Wu, and H. Xiong, “Towards controlling virus
propagation in information system with point-to-group information
sharing,” Decision Support Systems, vol. 48, no. 1, pp. 57–68, 2009.
[24] M.E.J.Newman, “Detecting community structure in networks,” The
European Physical Journal B - Condensed Matter and Complex
Systems, vol. 38, pp. 321–330, March 2004.
[25] S. Tanachaiwiwat and A. Helmy, “Encounter-based worms: Analysis and defense,” in in IEEE Conference on Sensor and Ad Hoc
Communications and Networks (SECON) Poster/Demo Session,
2006.
0.1
0.05
0.1
0.15
0.2
Initial ratio of sick persons
0.25
(c) vaccine: 22; alert: 18
0.3
0
0.05
0.1
0.15
0.2
Initial ratio of sick persons
0.25
0.3
(d) vaccine: 22; alert: unlimited
Fig. 8. Italian Traces: Performance comparison under different number
of vaccination and alert messages with pin = 0.5, pal = 0.7, pre =
0.1, pim = 0.2, pva = 0.3, Np = 4 days..
at which an infectious disease spreads in the healthcare
domain. The extracted social community information is
used for efficient vaccine distribution as opposed to the
traditional random vaccine distribution. Our framework
first partitions the set of encountered people into multiple
communities and kernel structures based on their social
relationships, where the people encountering information
can be derived from traces collected by mobile phones. We
believe people who are in the same kernel structure and
community as a sick person have higher risks of being
infected since they frequently interact with each other.
Hence, these people will be chosen by our framework
to receive vaccination or alert messages. We further developed a merging technique that helps to capture the
dynamic community information so as to control the
disease propagation more effectively. We compared our
community based disease control method with existing
techniques such as Random Distribution and Encounterbased methods using real contact-based traces such as
the MIT reality and Italian traces. Our results showed
that the propagation rate of an infectious disease can be
significantly reduced by utilizing the social community information. In addition, we compare our approach that utilizes dynamic community information with a betweennessbased approach using static community information. The
results also confirmed our observation that our community
based method is more effective in achieving a lower
infection ratio. Our study demonstrated more opportunities
for utilizing social relationships information to support
healthcare related applications.
R EFERENCES
[1] J.Scott, Social Network Analysis: A Handbook.
Ltd, 2000.
Sage Publication
10